RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-31 16:02:52 +01:00
Code Issues Projects Releases Wiki Activity
eb5d0cb4f4
llvm-mirror/test/CodeGen/X86/iv-users-in-other-loops.ll

297 lines
18 KiB
LLVM
Raw Normal View History

Eliminate more uses of llvm-as and llvm-dis. llvm-svn: 81290
2009-09-09 01:54:48 +02:00
; RUN: llc < %s -march=x86-64 -o %t
Reapply the new LoopStrengthReduction code, with compile time and bug fixes, and with improved heuristics for analyzing foreign-loop addrecs. This change also flattens IVUsers, eliminating the stride-oriented groupings, which makes it easier to work with. llvm-svn: 95975
2010-02-12 11:34:29 +01:00
; RUN: not grep inc %t
Reverse a loop that is counting up to a maximum to count down to 0 instead, under very restricted circumstances. Adjust 4 testcases in which this optimization fires. llvm-svn: 71439
2009-05-11 19:15:42 +02:00
; RUN: grep dec %t | count 2
Propagate the AlignStack bit in InlineAsm's to the PrologEpilog code, and use it to determine whether the asm forces stack alignment or not. gcc consistently does not do this for GCC-style asms; Apple gcc inconsistently sometimes does it for asm blocks. There is no convenient place to put a bit in either the SDNode or the MachineInstr form, so I've added an extra operand to each; unlovely, but it does allow for expansion for more bits, should we need it. PR 5125. Some existing testcases are affected. The operand lists of the SDNode and MachineInstr forms are indexed with awesome mnemonics, like "2"; I may fix this someday, but not now. I'm not making it any worse. If anyone is inspired I think you can find all the right places from this patch. llvm-svn: 107506
2010-07-02 22:16:09 +02:00
; RUN: grep addq %t | count 12
Remove the code from IVUsers that attempted to handle casted induction variables in cases where the cast isn't foldable. It ended up being a pessimization in many cases. This could be fixed, but it would require a bunch of complicated code in IVUsers' clients. The advantages of this approach aren't visible enough to justify it at this time. llvm-svn: 73706
2009-06-18 18:54:06 +02:00
; RUN: not grep addb %t
Reapply the new LoopStrengthReduction code, with compile time and bug fixes, and with improved heuristics for analyzing foreign-loop addrecs. This change also flattens IVUsers, eliminating the stride-oriented groupings, which makes it easier to work with. llvm-svn: 95975
2010-02-12 11:34:29 +01:00
; RUN: not grep leaq %t
; RUN: not grep leal %t
; RUN: not grep movq %t
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
; IV users in each of the loops from other loops shouldn't cause LSR
; to insert new induction variables. Previously it would create a
; flood of new induction variables.
Reverse a loop that is counting up to a maximum to count down to 0 instead, under very restricted circumstances. Adjust 4 testcases in which this optimization fires. llvm-svn: 71439
2009-05-11 19:15:42 +02:00
; Also, the loop reversal should kick in once.
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
target triple = "x86_64-unknown-linux-gnu"
define void @foo(float* %A, i32 %IA, float* %B, i32 %IB, float* nocapture %C, i32 %N) nounwind {
entry:
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%0 = xor i32 %IA, 1 ; <i32> [#uses=1]
%1 = xor i32 %IB, 1 ; <i32> [#uses=1]
%2 = or i32 %1, %0 ; <i32> [#uses=1]
%3 = icmp eq i32 %2, 0 ; <i1> [#uses=1]
br i1 %3, label %bb2, label %bb13
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
bb: ; preds = %bb3
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%4 = load float* %A_addr.0, align 4 ; <float> [#uses=1]
%5 = load float* %B_addr.0, align 4 ; <float> [#uses=1]
%6 = fmul float %4, %5 ; <float> [#uses=1]
%7 = fadd float %6, %Sum0.0 ; <float> [#uses=1]
%indvar.next154 = add i64 %B_addr.0.rec, 1 ; <i64> [#uses=1]
br label %bb2
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
bb2: ; preds = %entry, %bb
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%B_addr.0.rec = phi i64 [ %indvar.next154, %bb ], [ 0, %entry ] ; <i64> [#uses=14]
%Sum0.0 = phi float [ %7, %bb ], [ 0.000000e+00, %entry ] ; <float> [#uses=5]
%indvar146 = trunc i64 %B_addr.0.rec to i32 ; <i32> [#uses=1]
%N_addr.0 = sub i32 %N, %indvar146 ; <i32> [#uses=6]
%A_addr.0 = getelementptr float* %A, i64 %B_addr.0.rec ; <float*> [#uses=4]
%B_addr.0 = getelementptr float* %B, i64 %B_addr.0.rec ; <float*> [#uses=4]
%8 = icmp sgt i32 %N_addr.0, 0 ; <i1> [#uses=1]
br i1 %8, label %bb3, label %bb4
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
bb3: ; preds = %bb2
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%9 = ptrtoint float* %A_addr.0 to i64 ; <i64> [#uses=1]
%10 = and i64 %9, 15 ; <i64> [#uses=1]
%11 = icmp eq i64 %10, 0 ; <i1> [#uses=1]
br i1 %11, label %bb4, label %bb
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
bb4: ; preds = %bb3, %bb2
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%12 = ptrtoint float* %B_addr.0 to i64 ; <i64> [#uses=1]
%13 = and i64 %12, 15 ; <i64> [#uses=1]
%14 = icmp eq i64 %13, 0 ; <i1> [#uses=1]
%15 = icmp sgt i32 %N_addr.0, 15 ; <i1> [#uses=2]
br i1 %14, label %bb6.preheader, label %bb10.preheader
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
bb10.preheader: ; preds = %bb4
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
br i1 %15, label %bb9, label %bb12.loopexit
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
bb6.preheader: ; preds = %bb4
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
br i1 %15, label %bb5, label %bb8.loopexit
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
bb5: ; preds = %bb5, %bb6.preheader
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%indvar143 = phi i64 [ 0, %bb6.preheader ], [ %indvar.next144, %bb5 ] ; <i64> [#uses=3]
%vSum0.072 = phi <4 x float> [ zeroinitializer, %bb6.preheader ], [ %21, %bb5 ] ; <<4 x float>> [#uses=1]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%vSum1.070 = phi <4 x float> [ zeroinitializer, %bb6.preheader ], [ %29, %bb5 ] ; <<4 x float>> [#uses=1]
%vSum2.069 = phi <4 x float> [ zeroinitializer, %bb6.preheader ], [ %37, %bb5 ] ; <<4 x float>> [#uses=1]
%vSum3.067 = phi <4 x float> [ zeroinitializer, %bb6.preheader ], [ %45, %bb5 ] ; <<4 x float>> [#uses=1]
%indvar145 = trunc i64 %indvar143 to i32 ; <i32> [#uses=1]
%tmp150 = mul i32 %indvar145, -16 ; <i32> [#uses=1]
%N_addr.268 = add i32 %tmp150, %N_addr.0 ; <i32> [#uses=1]
%A_addr.273.rec = shl i64 %indvar143, 4 ; <i64> [#uses=5]
%B_addr.0.sum180 = add i64 %B_addr.0.rec, %A_addr.273.rec ; <i64> [#uses=2]
%B_addr.271 = getelementptr float* %B, i64 %B_addr.0.sum180 ; <float*> [#uses=1]
%A_addr.273 = getelementptr float* %A, i64 %B_addr.0.sum180 ; <float*> [#uses=1]
tail call void asm sideeffect ";# foo", "~{dirflag},~{fpsr},~{flags}"() nounwind
%16 = bitcast float* %A_addr.273 to <4 x float>* ; <<4 x float>*> [#uses=1]
%17 = load <4 x float>* %16, align 16 ; <<4 x float>> [#uses=1]
%18 = bitcast float* %B_addr.271 to <4 x float>* ; <<4 x float>*> [#uses=1]
%19 = load <4 x float>* %18, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%20 = fmul <4 x float> %17, %19 ; <<4 x float>> [#uses=1]
%21 = fadd <4 x float> %20, %vSum0.072 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%A_addr.273.sum163 = or i64 %A_addr.273.rec, 4 ; <i64> [#uses=1]
%A_addr.0.sum175 = add i64 %B_addr.0.rec, %A_addr.273.sum163 ; <i64> [#uses=2]
%22 = getelementptr float* %A, i64 %A_addr.0.sum175 ; <float*> [#uses=1]
%23 = bitcast float* %22 to <4 x float>* ; <<4 x float>*> [#uses=1]
%24 = load <4 x float>* %23, align 16 ; <<4 x float>> [#uses=1]
%25 = getelementptr float* %B, i64 %A_addr.0.sum175 ; <float*> [#uses=1]
%26 = bitcast float* %25 to <4 x float>* ; <<4 x float>*> [#uses=1]
%27 = load <4 x float>* %26, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%28 = fmul <4 x float> %24, %27 ; <<4 x float>> [#uses=1]
%29 = fadd <4 x float> %28, %vSum1.070 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%A_addr.273.sum161 = or i64 %A_addr.273.rec, 8 ; <i64> [#uses=1]
%A_addr.0.sum174 = add i64 %B_addr.0.rec, %A_addr.273.sum161 ; <i64> [#uses=2]
%30 = getelementptr float* %A, i64 %A_addr.0.sum174 ; <float*> [#uses=1]
%31 = bitcast float* %30 to <4 x float>* ; <<4 x float>*> [#uses=1]
%32 = load <4 x float>* %31, align 16 ; <<4 x float>> [#uses=1]
%33 = getelementptr float* %B, i64 %A_addr.0.sum174 ; <float*> [#uses=1]
%34 = bitcast float* %33 to <4 x float>* ; <<4 x float>*> [#uses=1]
%35 = load <4 x float>* %34, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%36 = fmul <4 x float> %32, %35 ; <<4 x float>> [#uses=1]
%37 = fadd <4 x float> %36, %vSum2.069 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%A_addr.273.sum159 = or i64 %A_addr.273.rec, 12 ; <i64> [#uses=1]
%A_addr.0.sum173 = add i64 %B_addr.0.rec, %A_addr.273.sum159 ; <i64> [#uses=2]
%38 = getelementptr float* %A, i64 %A_addr.0.sum173 ; <float*> [#uses=1]
%39 = bitcast float* %38 to <4 x float>* ; <<4 x float>*> [#uses=1]
%40 = load <4 x float>* %39, align 16 ; <<4 x float>> [#uses=1]
%41 = getelementptr float* %B, i64 %A_addr.0.sum173 ; <float*> [#uses=1]
%42 = bitcast float* %41 to <4 x float>* ; <<4 x float>*> [#uses=1]
%43 = load <4 x float>* %42, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%44 = fmul <4 x float> %40, %43 ; <<4 x float>> [#uses=1]
%45 = fadd <4 x float> %44, %vSum3.067 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%.rec83 = add i64 %A_addr.273.rec, 16 ; <i64> [#uses=1]
%A_addr.0.sum172 = add i64 %B_addr.0.rec, %.rec83 ; <i64> [#uses=2]
%46 = getelementptr float* %A, i64 %A_addr.0.sum172 ; <float*> [#uses=1]
%47 = getelementptr float* %B, i64 %A_addr.0.sum172 ; <float*> [#uses=1]
%48 = add i32 %N_addr.268, -16 ; <i32> [#uses=2]
%49 = icmp sgt i32 %48, 15 ; <i1> [#uses=1]
%indvar.next144 = add i64 %indvar143, 1 ; <i64> [#uses=1]
br i1 %49, label %bb5, label %bb8.loopexit
bb7: ; preds = %bb7, %bb8.loopexit
%indvar130 = phi i64 [ 0, %bb8.loopexit ], [ %indvar.next131, %bb7 ] ; <i64> [#uses=3]
%vSum0.260 = phi <4 x float> [ %vSum0.0.lcssa, %bb8.loopexit ], [ %55, %bb7 ] ; <<4 x float>> [#uses=1]
%indvar132 = trunc i64 %indvar130 to i32 ; <i32> [#uses=1]
%tmp133 = mul i32 %indvar132, -4 ; <i32> [#uses=1]
%N_addr.358 = add i32 %tmp133, %N_addr.2.lcssa ; <i32> [#uses=1]
%A_addr.361.rec = shl i64 %indvar130, 2 ; <i64> [#uses=3]
%B_addr.359 = getelementptr float* %B_addr.2.lcssa, i64 %A_addr.361.rec ; <float*> [#uses=1]
%A_addr.361 = getelementptr float* %A_addr.2.lcssa, i64 %A_addr.361.rec ; <float*> [#uses=1]
%50 = bitcast float* %A_addr.361 to <4 x float>* ; <<4 x float>*> [#uses=1]
%51 = load <4 x float>* %50, align 16 ; <<4 x float>> [#uses=1]
%52 = bitcast float* %B_addr.359 to <4 x float>* ; <<4 x float>*> [#uses=1]
%53 = load <4 x float>* %52, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%54 = fmul <4 x float> %51, %53 ; <<4 x float>> [#uses=1]
%55 = fadd <4 x float> %54, %vSum0.260 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%.rec85 = add i64 %A_addr.361.rec, 4 ; <i64> [#uses=2]
%56 = getelementptr float* %A_addr.2.lcssa, i64 %.rec85 ; <float*> [#uses=1]
%57 = getelementptr float* %B_addr.2.lcssa, i64 %.rec85 ; <float*> [#uses=1]
%58 = add i32 %N_addr.358, -4 ; <i32> [#uses=2]
%59 = icmp sgt i32 %58, 3 ; <i1> [#uses=1]
%indvar.next131 = add i64 %indvar130, 1 ; <i64> [#uses=1]
br i1 %59, label %bb7, label %bb13
bb8.loopexit: ; preds = %bb5, %bb6.preheader
%A_addr.2.lcssa = phi float* [ %A_addr.0, %bb6.preheader ], [ %46, %bb5 ] ; <float*> [#uses=3]
%vSum0.0.lcssa = phi <4 x float> [ zeroinitializer, %bb6.preheader ], [ %21, %bb5 ] ; <<4 x float>> [#uses=2]
%B_addr.2.lcssa = phi float* [ %B_addr.0, %bb6.preheader ], [ %47, %bb5 ] ; <float*> [#uses=3]
%vSum1.0.lcssa = phi <4 x float> [ zeroinitializer, %bb6.preheader ], [ %29, %bb5 ] ; <<4 x float>> [#uses=2]
%vSum2.0.lcssa = phi <4 x float> [ zeroinitializer, %bb6.preheader ], [ %37, %bb5 ] ; <<4 x float>> [#uses=2]
%N_addr.2.lcssa = phi i32 [ %N_addr.0, %bb6.preheader ], [ %48, %bb5 ] ; <i32> [#uses=3]
%vSum3.0.lcssa = phi <4 x float> [ zeroinitializer, %bb6.preheader ], [ %45, %bb5 ] ; <<4 x float>> [#uses=2]
%60 = icmp sgt i32 %N_addr.2.lcssa, 3 ; <i1> [#uses=1]
br i1 %60, label %bb7, label %bb13
bb9: ; preds = %bb9, %bb10.preheader
%indvar106 = phi i64 [ 0, %bb10.preheader ], [ %indvar.next107, %bb9 ] ; <i64> [#uses=3]
%vSum0.339 = phi <4 x float> [ zeroinitializer, %bb10.preheader ], [ %75, %bb9 ] ; <<4 x float>> [#uses=1]
%vSum1.237 = phi <4 x float> [ zeroinitializer, %bb10.preheader ], [ %80, %bb9 ] ; <<4 x float>> [#uses=1]
%vSum2.236 = phi <4 x float> [ zeroinitializer, %bb10.preheader ], [ %85, %bb9 ] ; <<4 x float>> [#uses=1]
%vSum3.234 = phi <4 x float> [ zeroinitializer, %bb10.preheader ], [ %90, %bb9 ] ; <<4 x float>> [#uses=1]
%indvar108 = trunc i64 %indvar106 to i32 ; <i32> [#uses=1]
%tmp113 = mul i32 %indvar108, -16 ; <i32> [#uses=1]
%N_addr.435 = add i32 %tmp113, %N_addr.0 ; <i32> [#uses=1]
%A_addr.440.rec = shl i64 %indvar106, 4 ; <i64> [#uses=5]
%B_addr.0.sum = add i64 %B_addr.0.rec, %A_addr.440.rec ; <i64> [#uses=2]
%B_addr.438 = getelementptr float* %B, i64 %B_addr.0.sum ; <float*> [#uses=1]
%A_addr.440 = getelementptr float* %A, i64 %B_addr.0.sum ; <float*> [#uses=1]
Change these tests to use regular loads instead of llvm.x86.sse2.loadu.dq. Enhance instcombine to use the preferred field of GetOrEnforceKnownAlignment in more cases, so that regular IR operations are optimized in the same way that the intrinsics currently are. llvm-svn: 64623
2009-02-16 01:44:23 +01:00
%61 = bitcast float* %B_addr.438 to <4 x float>* ; <i8*> [#uses=1]
%62 = load <4 x float>* %61, align 1
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%B_addr.438.sum169 = or i64 %A_addr.440.rec, 4 ; <i64> [#uses=1]
%B_addr.0.sum187 = add i64 %B_addr.0.rec, %B_addr.438.sum169 ; <i64> [#uses=2]
%63 = getelementptr float* %B, i64 %B_addr.0.sum187 ; <float*> [#uses=1]
Change these tests to use regular loads instead of llvm.x86.sse2.loadu.dq. Enhance instcombine to use the preferred field of GetOrEnforceKnownAlignment in more cases, so that regular IR operations are optimized in the same way that the intrinsics currently are. llvm-svn: 64623
2009-02-16 01:44:23 +01:00
%64 = bitcast float* %63 to <4 x float>* ; <i8*> [#uses=1]
%65 = load <4 x float>* %64, align 1
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%B_addr.438.sum168 = or i64 %A_addr.440.rec, 8 ; <i64> [#uses=1]
%B_addr.0.sum186 = add i64 %B_addr.0.rec, %B_addr.438.sum168 ; <i64> [#uses=2]
%66 = getelementptr float* %B, i64 %B_addr.0.sum186 ; <float*> [#uses=1]
Change these tests to use regular loads instead of llvm.x86.sse2.loadu.dq. Enhance instcombine to use the preferred field of GetOrEnforceKnownAlignment in more cases, so that regular IR operations are optimized in the same way that the intrinsics currently are. llvm-svn: 64623
2009-02-16 01:44:23 +01:00
%67 = bitcast float* %66 to <4 x float>* ; <i8*> [#uses=1]
%68 = load <4 x float>* %67, align 1
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%B_addr.438.sum167 = or i64 %A_addr.440.rec, 12 ; <i64> [#uses=1]
%B_addr.0.sum185 = add i64 %B_addr.0.rec, %B_addr.438.sum167 ; <i64> [#uses=2]
%69 = getelementptr float* %B, i64 %B_addr.0.sum185 ; <float*> [#uses=1]
Change these tests to use regular loads instead of llvm.x86.sse2.loadu.dq. Enhance instcombine to use the preferred field of GetOrEnforceKnownAlignment in more cases, so that regular IR operations are optimized in the same way that the intrinsics currently are. llvm-svn: 64623
2009-02-16 01:44:23 +01:00
%70 = bitcast float* %69 to <4 x float>* ; <i8*> [#uses=1]
%71 = load <4 x float>* %70, align 1
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%72 = bitcast float* %A_addr.440 to <4 x float>* ; <<4 x float>*> [#uses=1]
%73 = load <4 x float>* %72, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%74 = fmul <4 x float> %73, %62 ; <<4 x float>> [#uses=1]
%75 = fadd <4 x float> %74, %vSum0.339 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%76 = getelementptr float* %A, i64 %B_addr.0.sum187 ; <float*> [#uses=1]
%77 = bitcast float* %76 to <4 x float>* ; <<4 x float>*> [#uses=1]
%78 = load <4 x float>* %77, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%79 = fmul <4 x float> %78, %65 ; <<4 x float>> [#uses=1]
%80 = fadd <4 x float> %79, %vSum1.237 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%81 = getelementptr float* %A, i64 %B_addr.0.sum186 ; <float*> [#uses=1]
%82 = bitcast float* %81 to <4 x float>* ; <<4 x float>*> [#uses=1]
%83 = load <4 x float>* %82, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%84 = fmul <4 x float> %83, %68 ; <<4 x float>> [#uses=1]
%85 = fadd <4 x float> %84, %vSum2.236 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%86 = getelementptr float* %A, i64 %B_addr.0.sum185 ; <float*> [#uses=1]
%87 = bitcast float* %86 to <4 x float>* ; <<4 x float>*> [#uses=1]
%88 = load <4 x float>* %87, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%89 = fmul <4 x float> %88, %71 ; <<4 x float>> [#uses=1]
%90 = fadd <4 x float> %89, %vSum3.234 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%.rec89 = add i64 %A_addr.440.rec, 16 ; <i64> [#uses=1]
%A_addr.0.sum170 = add i64 %B_addr.0.rec, %.rec89 ; <i64> [#uses=2]
%91 = getelementptr float* %A, i64 %A_addr.0.sum170 ; <float*> [#uses=1]
%92 = getelementptr float* %B, i64 %A_addr.0.sum170 ; <float*> [#uses=1]
%93 = add i32 %N_addr.435, -16 ; <i32> [#uses=2]
%94 = icmp sgt i32 %93, 15 ; <i1> [#uses=1]
%indvar.next107 = add i64 %indvar106, 1 ; <i64> [#uses=1]
br i1 %94, label %bb9, label %bb12.loopexit
bb11: ; preds = %bb11, %bb12.loopexit
%indvar = phi i64 [ 0, %bb12.loopexit ], [ %indvar.next, %bb11 ] ; <i64> [#uses=3]
%vSum0.428 = phi <4 x float> [ %vSum0.3.lcssa, %bb12.loopexit ], [ %100, %bb11 ] ; <<4 x float>> [#uses=1]
%indvar96 = trunc i64 %indvar to i32 ; <i32> [#uses=1]
%tmp = mul i32 %indvar96, -4 ; <i32> [#uses=1]
%N_addr.526 = add i32 %tmp, %N_addr.4.lcssa ; <i32> [#uses=1]
%A_addr.529.rec = shl i64 %indvar, 2 ; <i64> [#uses=3]
%B_addr.527 = getelementptr float* %B_addr.4.lcssa, i64 %A_addr.529.rec ; <float*> [#uses=1]
%A_addr.529 = getelementptr float* %A_addr.4.lcssa, i64 %A_addr.529.rec ; <float*> [#uses=1]
Change these tests to use regular loads instead of llvm.x86.sse2.loadu.dq. Enhance instcombine to use the preferred field of GetOrEnforceKnownAlignment in more cases, so that regular IR operations are optimized in the same way that the intrinsics currently are. llvm-svn: 64623
2009-02-16 01:44:23 +01:00
%95 = bitcast float* %B_addr.527 to <4 x float>* ; <i8*> [#uses=1]
%96 = load <4 x float>* %95, align 1
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%97 = bitcast float* %A_addr.529 to <4 x float>* ; <<4 x float>*> [#uses=1]
%98 = load <4 x float>* %97, align 16 ; <<4 x float>> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%99 = fmul <4 x float> %98, %96 ; <<4 x float>> [#uses=1]
%100 = fadd <4 x float> %99, %vSum0.428 ; <<4 x float>> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%.rec91 = add i64 %A_addr.529.rec, 4 ; <i64> [#uses=2]
%101 = getelementptr float* %A_addr.4.lcssa, i64 %.rec91 ; <float*> [#uses=1]
%102 = getelementptr float* %B_addr.4.lcssa, i64 %.rec91 ; <float*> [#uses=1]
%103 = add i32 %N_addr.526, -4 ; <i32> [#uses=2]
%104 = icmp sgt i32 %103, 3 ; <i1> [#uses=1]
%indvar.next = add i64 %indvar, 1 ; <i64> [#uses=1]
br i1 %104, label %bb11, label %bb13
bb12.loopexit: ; preds = %bb9, %bb10.preheader
%A_addr.4.lcssa = phi float* [ %A_addr.0, %bb10.preheader ], [ %91, %bb9 ] ; <float*> [#uses=3]
%vSum0.3.lcssa = phi <4 x float> [ zeroinitializer, %bb10.preheader ], [ %75, %bb9 ] ; <<4 x float>> [#uses=2]
%B_addr.4.lcssa = phi float* [ %B_addr.0, %bb10.preheader ], [ %92, %bb9 ] ; <float*> [#uses=3]
%vSum1.2.lcssa = phi <4 x float> [ zeroinitializer, %bb10.preheader ], [ %80, %bb9 ] ; <<4 x float>> [#uses=2]
%vSum2.2.lcssa = phi <4 x float> [ zeroinitializer, %bb10.preheader ], [ %85, %bb9 ] ; <<4 x float>> [#uses=2]
%N_addr.4.lcssa = phi i32 [ %N_addr.0, %bb10.preheader ], [ %93, %bb9 ] ; <i32> [#uses=3]
%vSum3.2.lcssa = phi <4 x float> [ zeroinitializer, %bb10.preheader ], [ %90, %bb9 ] ; <<4 x float>> [#uses=2]
%105 = icmp sgt i32 %N_addr.4.lcssa, 3 ; <i1> [#uses=1]
br i1 %105, label %bb11, label %bb13
bb13: ; preds = %bb12.loopexit, %bb11, %bb8.loopexit, %bb7, %entry
%Sum0.1 = phi float [ 0.000000e+00, %entry ], [ %Sum0.0, %bb7 ], [ %Sum0.0, %bb8.loopexit ], [ %Sum0.0, %bb11 ], [ %Sum0.0, %bb12.loopexit ] ; <float> [#uses=1]
%vSum3.1 = phi <4 x float> [ zeroinitializer, %entry ], [ %vSum3.0.lcssa, %bb7 ], [ %vSum3.0.lcssa, %bb8.loopexit ], [ %vSum3.2.lcssa, %bb11 ], [ %vSum3.2.lcssa, %bb12.loopexit ] ; <<4 x float>> [#uses=1]
%N_addr.1 = phi i32 [ %N, %entry ], [ %N_addr.2.lcssa, %bb8.loopexit ], [ %58, %bb7 ], [ %N_addr.4.lcssa, %bb12.loopexit ], [ %103, %bb11 ] ; <i32> [#uses=2]
%vSum2.1 = phi <4 x float> [ zeroinitializer, %entry ], [ %vSum2.0.lcssa, %bb7 ], [ %vSum2.0.lcssa, %bb8.loopexit ], [ %vSum2.2.lcssa, %bb11 ], [ %vSum2.2.lcssa, %bb12.loopexit ] ; <<4 x float>> [#uses=1]
%vSum1.1 = phi <4 x float> [ zeroinitializer, %entry ], [ %vSum1.0.lcssa, %bb7 ], [ %vSum1.0.lcssa, %bb8.loopexit ], [ %vSum1.2.lcssa, %bb11 ], [ %vSum1.2.lcssa, %bb12.loopexit ] ; <<4 x float>> [#uses=1]
%B_addr.1 = phi float* [ %B, %entry ], [ %B_addr.2.lcssa, %bb8.loopexit ], [ %57, %bb7 ], [ %B_addr.4.lcssa, %bb12.loopexit ], [ %102, %bb11 ] ; <float*> [#uses=1]
%vSum0.1 = phi <4 x float> [ zeroinitializer, %entry ], [ %vSum0.0.lcssa, %bb8.loopexit ], [ %55, %bb7 ], [ %vSum0.3.lcssa, %bb12.loopexit ], [ %100, %bb11 ] ; <<4 x float>> [#uses=1]
%A_addr.1 = phi float* [ %A, %entry ], [ %A_addr.2.lcssa, %bb8.loopexit ], [ %56, %bb7 ], [ %A_addr.4.lcssa, %bb12.loopexit ], [ %101, %bb11 ] ; <float*> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%106 = fadd <4 x float> %vSum0.1, %vSum2.1 ; <<4 x float>> [#uses=1]
%107 = fadd <4 x float> %vSum1.1, %vSum3.1 ; <<4 x float>> [#uses=1]
%108 = fadd <4 x float> %106, %107 ; <<4 x float>> [#uses=4]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%tmp23 = extractelement <4 x float> %108, i32 0 ; <float> [#uses=1]
%tmp21 = extractelement <4 x float> %108, i32 1 ; <float> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%109 = fadd float %tmp23, %tmp21 ; <float> [#uses=1]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%tmp19 = extractelement <4 x float> %108, i32 2 ; <float> [#uses=1]
%tmp17 = extractelement <4 x float> %108, i32 3 ; <float> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%110 = fadd float %tmp19, %tmp17 ; <float> [#uses=1]
%111 = fadd float %109, %110 ; <float> [#uses=1]
%Sum0.254 = fadd float %111, %Sum0.1 ; <float> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%112 = icmp sgt i32 %N_addr.1, 0 ; <i1> [#uses=1]
br i1 %112, label %bb.nph56, label %bb16
bb.nph56: ; preds = %bb13
%tmp. = zext i32 %N_addr.1 to i64 ; <i64> [#uses=1]
br label %bb14
bb14: ; preds = %bb14, %bb.nph56
%indvar117 = phi i64 [ 0, %bb.nph56 ], [ %indvar.next118, %bb14 ] ; <i64> [#uses=3]
%Sum0.255 = phi float [ %Sum0.254, %bb.nph56 ], [ %Sum0.2, %bb14 ] ; <float> [#uses=1]
%tmp.122 = sext i32 %IB to i64 ; <i64> [#uses=1]
%B_addr.652.rec = mul i64 %indvar117, %tmp.122 ; <i64> [#uses=1]
%tmp.124 = sext i32 %IA to i64 ; <i64> [#uses=1]
%A_addr.653.rec = mul i64 %indvar117, %tmp.124 ; <i64> [#uses=1]
%B_addr.652 = getelementptr float* %B_addr.1, i64 %B_addr.652.rec ; <float*> [#uses=1]
%A_addr.653 = getelementptr float* %A_addr.1, i64 %A_addr.653.rec ; <float*> [#uses=1]
%113 = load float* %A_addr.653, align 4 ; <float> [#uses=1]
%114 = load float* %B_addr.652, align 4 ; <float> [#uses=1]
Split the Add, Sub, and Mul instruction opcodes into separate integer and floating-point opcodes, introducing FAdd, FSub, and FMul. For now, the AsmParser, BitcodeReader, and IRBuilder all preserve backwards compatability, and the Core LLVM APIs preserve backwards compatibility for IR producers. Most front-ends won't need to change immediately. This implements the first step of the plan outlined here: http://nondot.org/sabre/LLVMNotes/IntegerOverflow.txt llvm-svn: 72897
2009-06-05 00:49:04 +02:00
%115 = fmul float %113, %114 ; <float> [#uses=1]
%Sum0.2 = fadd float %115, %Sum0.255 ; <float> [#uses=2]
Fix the code that checked if a SCEVAddRecExpr Start contains an addrec in a different loop to check the value being added to the accumulated Start value, not the Start value before it has the new value added to it. This prevents LSR from going crazy on the included testcase. Dale, please review. llvm-svn: 64440
2009-02-13 04:58:31 +01:00
%indvar.next118 = add i64 %indvar117, 1 ; <i64> [#uses=2]
%exitcond = icmp eq i64 %indvar.next118, %tmp. ; <i1> [#uses=1]
br i1 %exitcond, label %bb16, label %bb14
bb16: ; preds = %bb14, %bb13
%Sum0.2.lcssa = phi float [ %Sum0.254, %bb13 ], [ %Sum0.2, %bb14 ] ; <float> [#uses=1]
store float %Sum0.2.lcssa, float* %C, align 4
ret void
}
Reference in New Issue Copy Permalink